CincyDeliver 2026 · Behind the Talk

I brought the argument. AI helped me test it.

The talk is mine — the hypothesis, the judgment, the call on what ships. This is a plain account of where AI worked alongside me, and where it didn't.

The Transformation We Never Finished — a talk by Matt Anderson, VaporPoint. Sterling Room · Friday, July 24.

Matt led AI supported

The talk started with a hypothesis, not a prompt.

The thesis came first — you can't AI your way out of an organization that couldn't Agile — out of years inside transformations, the ones that took and the ones that didn't. I didn't ask a model to invent an argument. I brought one I already believed and needed to pressure-test.

The collaborator was Claude Code, with OpenAI's Codex pulled in as an independent checker and Firecrawl for web research. Across the build it acted as a research partner that chased down evidence in parallel, a skeptic that argued with me instead of agreeing, an editorial assistant for my ideas, and a builder that turned a fact-checked paper into a delivery package. It did not decide what the talk should say. That stayed with me.

AI was a collaborator here, not a crutch. Below is who did what, stage by stage. Each item is traceable to a working file — the process left its own record.


Five stages. My judgment on the left, AI's work on the right.

01

The hypothesis

Matt

Owned the core claim and the governing premise — that this is a thinking pattern, not a framework, drawn from real transformations rather than a methodology. The spine of the talk is a conviction I held before any tool touched it.

AI

Interrogated the idea one branch at a time — where it was sharp, where it was soft — and helped me shape a loose conviction into a structured argument with a clear spine.

Claude Code · structured scoping
02

The evidence

Matt

Decided which claims were load-bearing and had to be defensible — and which ideas were mine to stand behind regardless of what the data said.

AI

Ran two parallel research passes against 2026 primary sources: one to confirm, and a second deliberately built to disconfirm — sent to hunt for evidence that would prove me wrong and to retire stale statistics.

Claude Code · Firecrawl · confirming + adversarial passes
03

The stress test

Matt

Set the bar: the thesis only survives if the evidence does. I wanted the one falsifiable claim in the talk — the "kill-switch" — attacked, not protected.

AI

Three agents fact-checked the whole evidence base six days out and produced a diff-verified change ledger — six corrections, five precision fixes, two flagged stats. A second, independent model double-checked a key analytical mapping. The core thesis held.

Claude Code (3 agents) · OpenAI Codex cross-check
04

The craft

Matt

Chose the visual language, decided what each signature diagram had to mean, and did a final hand-markup pass on the printed deck.

AI

Rendered a 38-slide branded deck and the take-home field guide, screenshot-checking every slide for layout bugs, and organizing the fact-checked paper into a handout — compression, not new claims.

Claude Code · HTML/CSS render · headless-Chrome QA
05

The decision

Matt

Kept every call that matters: what makes the cut, what stays honest, and when it's ready to go in front of a room. AI drafts; I decide.

AI

Held its work as proposals — deck, handout, and publish, all waiting on my go-ahead. Nothing outward-facing went out without my sign-off.

principle · the collaborator proposes, the human commits

Four habits kept it honest.

A crutch generates and you accept it. Collaboration means building in ways for the tool to tell you you're wrong. Four did that here.

Argue, don't agree

A skeptic by design

One of the two Firecrawl research passes existed only to disconfirm — to find the evidence that would sink the argument, and to retire stale numbers.

Check the checker

More than one model

OpenAI's Codex independently checked a key mapping that Claude produced. Using one model to audit another is the same posture the talk argues for.

Show the receipts

Every claim is traceable

The fact-check produced a diff-verified change ledger — six corrections, five precision fixes, two flagged stats — and every number traces back to a working file and a 2026 source.

Drafts, never decides

The human commits

Deck, handout, and publish all waited on my go-ahead. The tool proposed; nothing outward-facing went live without my sign-off — the collaborator never got the last word.


AI didn't write my talk. It made me a faster researcher, a harder skeptic, and a better editor of my own ideas — and then it got out of the way so the argument could be mine.

— Matt Anderson, VaporPoint

The same lesson the talk is about.

You can't AI your way out of an organization that couldn't Agile — and you can't AI your way to a talk you don't already have. The tool amplified what I brought. It couldn't manufacture conviction, and I never asked it to.

If that's how one conference talk gets built, it's worth asking what it means for the transformations we're all working inside. That's the conversation on Friday.